Distinguish attempted from discarded function calls in RL training

Determine whether the zero tool executions in the Qwen2.5-Coder function-calling training arm arose because the model never attempted to call run_tests or because attempted calls were discarded by the interface before execution.

Background

The training instrumentation recorded parser-accepted and executed calls, but raw function-calling rollout text was not retained. Consequently, the training data cannot distinguish a failure to generate any tool-call attempt from generation of calls that the parser rejected. A separate probe at the same scale found no output naming run_tests, but that probe was not the original training run and therefore does not resolve the issue.